fine-tune

[JuShuai-gif]

Is it possible to fine-tune a pretrained pi05 model for a single-arm setup by modifying only the configuration file?

Specifically, the original pi05 configuration uses three image inputs and a 32-dimensional state vector as model inputs. In my case, I would like to fine-tune the model with two image inputs and a 7-dimensional state.

Can this be achieved solely by updating the model configuration, or are additional architectural changes (e.g., reinitializing certain projection layers or heads) required?

[Sakits]

Hi @JuShuai-gif ,

Yes, this can be done by only modifying the config.

During fine-tuning, our framework automatically reads the configuration from the LeRobot dataset, so your setup with 2 camera inputs and a 7-dimensional state will be applied automatically. You can directly run the fine-tuning command without making any changes, other than filling in the dataset path.

The only place where you need to specify the cameras is during inference, where you define the ordering and identity of the two cameras. :)

Best,
Jiaming